-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_mnist.py
executable file
·261 lines (221 loc) · 9.22 KB
/
train_mnist.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
import argparse
import copy
import numpy as np
import os
import time
import chainer
from chainer.dataset import convert
from chainer import functions as F
from chainer import serializers
import cupy
import matplotlib.pyplot as plt
from matplotlib.ticker import AutoMinorLocator
import seaborn as sns
from logging import getLogger, DEBUG, Formatter, StreamHandler
logger = getLogger(__name__)
handler = StreamHandler()
handler.setLevel(DEBUG)
logger.setLevel(DEBUG)
# handler.setFormatter(Formatter('%(asctime)s %(message)s'))
logger.addHandler(handler)
import nets
def visualize(result, name):
# (N_samples=5, timesteps=200, types)
result = np.array(result)
assert result.shape[2] == 2
# ax = sns.tsplot(data=result, condition=['OptNet', 'Adam'], linestyle='--')
def trans(series):
# (N_samples, timesteps)
x = np.tile(np.arange(series.shape[1]) + 1,
(series.shape[0], 1)).flatten()
y = series.flatten()
return {'x': x, 'y': y}
ax = sns.lineplot(label='OptNet', **trans(result[:, :, 0]))
ax = sns.lineplot(label='Adam', ax=ax, **trans(result[:, :, 1]))
ax.lines[-1].set_linestyle('-')
ax.legend()
plt.yscale('log'), plt.xlabel('steps')
plt.ylabel('loss'), plt.title('MNIST')
plt.ylim(0.09, 3.0)
plt.xlim(1, result.shape[1])
plt.grid(which='both', alpha=0.6, color='black', linewidth=0.1,
linestyle='-')
ax.tick_params(which='both', direction='in')
ax.tick_params(which='major', length=8)
ax.tick_params(which='minor', length=3)
ax.xaxis.set_minor_locator(AutoMinorLocator(5))
plt.show()
plt.savefig(name)
plt.close()
def evaluate_optimizer(args, train, optimizer):
if isinstance(optimizer, nets.optnets.LSTMOptNet):
optimizer.release_all()
device = chainer.get_device(args.gpu)
device.use()
n_evaluation_runs = args.evaluation_runs # 5?
max_iter_of_meta = args.iter_meta # 100
all_losses = []
for _ in range(n_evaluation_runs):
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
losses = []
model = nets.images.MLPforMNIST()
model.to_device(device)
optimizer.setup(model)
iteration = 0
while iteration < max_iter_of_meta:
# routine
iteration += 1
batch = train_iter.next()
batch = convert.concat_examples(batch, device=device)
x, t = batch
with chainer.using_config('train', True):
loss, acc = model(x, t, get_accuracy=True)
model.cleargrads()
loss.backward(retain_grad=False) # False
optimizer.update(train_optnet=False)
losses.append(loss.item()) # log
all_losses.append(losses)
# TODO: use losses in only last half iterations?
last10_mean = np.mean([losses[-10:] for losses in all_losses])
return last10_mean, all_losses
def pretraining(optimizer):
logger.info('pretraining')
copy_grand_opt = copy.deepcopy(optimizer.grand_optimizer)
losses = []
for _ in range(10):
x = optimizer.optnet.xp.random.normal(
scale=10., size=(10000, 1)).astype('f')
g = optimizer.optnet.step(x)
# loss forcing g's sign to be the flip of input's sign
# theta = theta - c*gradient
# theta = theta + g
loss = F.mean(F.clip(g, 0, 100) * (x > 0)
+ F.clip(-g, 0, 100) * (x < 0))
optimizer.optnet.cleargrads()
loss.backward()
optimizer.meta_update()
optimizer.optnet.reset_state()
losses.append(loss.item())
logger.info('finished pretraining. losses {}'.format(losses))
optimizer.release_all()
# reset adam state
optimizer = nets.optnets.OptimizerByNet(optimizer.optnet, copy_grand_opt)
return optimizer, copy_grand_opt
def set_seed(seed):
np.random.seed(seed)
cupy.random.seed(seed)
os.environ['CHAINER_SEED'] = str(seed)
logger.info('set seed {}'.format(seed))
def get_adam_result(args, test_data):
optimizer = chainer.optimizers.Adam(alpha=args.adam_alpha)
test_loss, test_all_losses = evaluate_optimizer(args, test_data, optimizer)
return test_loss, test_all_losses
def train_optimizer(args):
device = chainer.get_device(args.gpu)
device.use()
# prepare test (meta-inference) set from training data
train, _ = chainer.datasets.get_mnist(ndim=1)
train, test = chainer.datasets.split_dataset_random(
train, int(len(train) * 0.5), seed=2400)
# get results using adam
set_seed(args.seed)
adam_test_loss, adam_test_all_losses = get_adam_result(args, test)
if args.use_adam:
logger.info('use Adam')
logger.info('\t'*2 + 'TEST last10mean loss {:.5f}'.format(test_loss))
return None # finish
# make learnable optimizer and its optmizer
set_seed(args.seed)
logger.info('use MetaOpt')
optnet = nets.optnets.LSTMOptNet(out_scale=args.out_scale,
do_preprocess=True)
optnet.to_device(device)
grand_opt = chainer.optimizers.Adam(alpha=args.adam_alpha)
optimizer = nets.optnets.OptimizerByNet(optnet, grand_opt)
# hyperparams of training
max_cycle_of_meta = args.cycle_meta # many
max_iter_of_meta = args.iter_meta # 100
unroll_iters = args.unroll_iters # 20
optimizer_path = args.optimizer_path
best_test_loss = 100000000000 # inf
# original pretraining:
# fix bad initialization with anti-update
# where a model is updated to the direction of gradient.
# (typically, like SGD, it should be the opposite of gradient.)
if args.do_pretraining:
set_seed(args.seed)
optimizer, grand_opt = pretraining(optimizer)
# train
set_seed(args.seed)
logger.info('training start')
for i_cycle in range(max_cycle_of_meta):
# in each cycle,
# meta-train optimizer through training newly initialized model
time_cycle_start = time.time()
meta_losses, loss_logs, acc_logs = [], [], []
# init model
model = nets.images.MLPforMNIST()
model.to_device(device)
optimizer.setup(model)
# reset data iterator
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
iteration = 0
while iteration < max_iter_of_meta:
iteration += 1
batch = train_iter.next()
batch = convert.concat_examples(batch, device=device)
x, t = batch
with chainer.using_config('train', True):
loss, acc = model(x, t, get_accuracy=True)
model.cleargrads()
loss.backward(retain_grad=True)
optimizer.update()
meta_losses.append(loss) # stored for meta update
loss_logs.append(loss.item()) # log
acc_logs.append(acc) # log
# meta update
if len(meta_losses) >= unroll_iters:
optimizer.optnet.cleargrads()
sum(meta_losses).backward(retain_grad=True)
model.cleargrads()
optimizer.meta_update()
meta_losses = []
iter_per_sec = iteration / (time.time() - time_cycle_start)
logger.info(' cycle {}\tlast10mean loss {:.5f}\tacc {:.5f}\t({:.2f} i/s)'
.format(i_cycle, np.mean(loss_logs[-10:]), np.mean(acc_logs[-10:]), iter_per_sec))
test_loss, test_all_losses = evaluate_optimizer(
args, test, copy.deepcopy(optimizer))
logger.info('\t'*2 + 'TEST last10mean loss {:.5f}'.format(test_loss))
if test_loss < best_test_loss:
logger.info('save optimizer test_loss {:.4f} -> {:.4f}'
.format(best_test_loss, test_loss))
visualize(np.stack([test_all_losses, adam_test_all_losses], axis=2),
name=optimizer_path + '.testloss.png')
chainer.serializers.save_npz(optimizer_path, optimizer.optnet)
best_test_loss = test_loss
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--batchsize', '-b', type=int, default=128)
parser.add_argument('--gpu', '-g', type=int, default=0,
help='if cpu, use -1')
parser.add_argument('--out', '-o', default='result')
parser.add_argument('--cycle-meta', type=int, default=200) # unk
parser.add_argument('--evaluation-runs', type=int, default=5)
parser.add_argument('--iter-meta', type=int, default=100)
parser.add_argument('--unroll-iters', type=int, default=20)
parser.add_argument('--adam-alpha', type=float, default=0.03)
parser.add_argument('--out-scale', type=float, default=0.1)
parser.add_argument('--optimizer-path', type=str,
default='optnet_seed{seed}.npz')
parser.add_argument('--seed', type=int, default=7772)
parser.add_argument('--use-adam', action='store_true', default=False)
parser.add_argument('--do-pretraining', action='store_true', default=False)
args = parser.parse_args()
if '{seed}' in args.optimizer_path:
changed = args.optimizer_path.replace('{seed}', str(args.seed))
logger.info('optimizer_path {} -> {}'
.format(args.optimizer_path, changed))
args.optimizer_path = changed
train_optimizer(args)
if __name__ == '__main__':
main()